Frontier papers

U-Net:少样本图像分割的基线

U-Net: Convolutional Networks for Biomedical Image Segmentation

Ronneberger, Fischer, Brox · 2015 · arXiv:1505.04597

Foundation paper · permanent
2015
The question

在标注极少的情况下,如何让像素级分割既准又稳?

What they did

编码—解码 + 跳跃连接,把高分辨率细节直接接到上采样路径。

Evidence

多届 ISBI 细胞追踪挑战赛;小样本训练下仍能收敛。

Do not over-trust

换成灰度超声图,跳跃连接会把探头伪影也带到输出;跨批次要先看召回而非 IoU。

Limits

需要像素级标注;对低对比度、强噪声的工业图像,跨批次泛化未经验证。

In-class task

用 200 张自标注缺陷图训练,留一个批次做测试,报告跨批次召回与误检。

A foundation paper is the landmark work for a method and stays permanently: it explains why the method works, and is never replaced for being old.
Paper Related chapters: 标签噪声:数据质量决定上限 小样本:20 张图也能训 Back to learning